Ming-3.0-tts-tiny vs Sesame-Preview: A Checkpoint You Can Download Against a Voice You Cannot Call

Put Ming-3.0-tts-tiny and Sesame Preview in the same sentence and the comparison falls apart before the first benchmark, because only one of them is something a developer can obtain. Ming-3.0-tts-tiny is Ant Group's inclusionAI checkpoint: 7.9 billion parameters total, about 1.3 billion active per token, MIT-licensed weights published on Hugging Face with inference code that arrived on GitHub on October 19, 2026. Sesame Preview is a consumer assistant — four voices, an iOS and Android app, no API, no model ID, no published score for the voice that makes it worth writing about. One is a file. The other is a product. The interesting question is not which sounds better, since nobody has measured that on either side in a way that can be compared, but what each one lets you do on a Tuesday afternoon when you have a script and a deadline. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

The dating matters here in a way it rarely does. Ming-3.0-tts-tiny's repository was created on September 22, 2026, the weights went up on September 24, and for three and a half weeks the release was unusable — a folder of shards with no published way to run them. The inference code, the one-click examples and the task cookbook landed together on October 19 at 08:58 UTC. Ant has not announced any of it. Sesame Preview, by contrast, has been shipping in public since the spring and got a real update on October 6, 2026, when the company's chief product officer described a new Conversational Speech Model behind the app. Neither of those is a launch in the sense a press release means, but both are movements, and only one of them comes with something you can download.

The short version, before the evidence

If you are building a product, Sesame Preview is not a candidate and cannot become one today: there is no endpoint to call, no licence to read, no price, and no commitment that the voice in the app is the voice in any model. If you are building a product and need a voice you own, Ming-3.0-tts-tiny is a candidate — a young, unscored, self-hosted one that will make you do your own evaluation.

That is the whole decision for most teams. Everything below is for the cases where it is not: the ones comparing them as conversational audio rather than as products, or weighing whether Sesame's approach is the one to copy.

What each one actually puts in your hands

• What you download — Ming-3.0-tts-tiny: roughly 18 GB containing the Ling V3 backbone, a diffusion-transformer head, an audio tokenizer, a CAM++ speaker encoder and a text tokenizer. Sesame Preview: an app from the App Store or G​oogle Play, and nothing else.

• What you can call — Ming-3.0-tts-tiny: a Python class, MingTinyTTS.from_pretrained, with task names for zero-shot cloning, multilingual synthesis, paralinguistics, podcast dialogue, emotion, dialect, free-form style and voice design. Sesame: a chat interface with four personalities. No model ID has ever been published for the production voice.

• Licence — Ming-3.0-tts-tiny: MIT on weights and code, with a usage note asking you not to impersonate people and to label synthetic audio. Sesame: the app's terms of service, plus an open checkpoint discussed below that is not the production model.

• Language coverage — Ming-3.0-tts-tiny: Chinese, English, German, Spanish, French, Korean and Japanese, with Sichuanese and Cantonese available from a dialect reference clip. Sesame: English is the only officially supported language; the guide notes the assistants also handle French, Italian, German, Spanish, Chinese, Japanese and Korean "early", which reads as a warning rather than a feature list.

• Price — Ming-3.0-tts-tiny: no rate card; you pay for an H800 and the electricity. Sesame: free to use, no published plan tier, and the company says it runs no ads and does not sell data.

• Independent score — Ming-3.0-tts-tiny: none, and specifically not the 1,188 Elo that Artificial Analysis lists for "Ming-3.0-tts", which belongs to the 0.5B checkpoint. Sesame: none for the app's voice either. This is the rare matchup where neither side has a number.

The one Sesame artefact a developer can hold

There is a Sesame model on Hugging Face, and it is worth being precise about what it is, because it is routinely mistaken for the voice in the app.

CSM-1B is Apache-2.0, English-only, and built on the architecture the company has described publicly: a Llama backbone predicting the first audio codebook and a smaller depth decoder predicting the rest, decoding through Kyutai's Mimi codec at 24 kHz. The card calls it "a base generation model" — capable of producing varied voices, not tailored to any particular one, and explicitly a derivative of it is what powers the interactive demo rather than the checkpoint being the demo. As of this capture it showed about 130,000 downloads in the previous month and roughly 2,500 likes, which is a real developer following for a two-year-old research release.

What that means in practice: you can load Sesame's architecture, fine-tune it, and self-host a conversational T​TS. You cannot load Sesame's product. The October update — a new CSM with better expressivity and pronunciation, in the company's own words — moved the thing you cannot have, and said nothing measurable about it.

Where they genuinely overlap, and where the overlap ends

Both are aimed at speech that sounds like conversation rather than narration, and that is not a small thing to have in common. Sesame built its reputation on breath, hesitation and interruption timing so convincing that early listeners assumed a person was on the line. Ant's tiny model ships 14 inline paralinguistic tags for laughter, sighs, breath and hesitation, and a multi-speaker podcast task that its own 0.5B sibling cannot do.

Past that, the overlap thins quickly.

Sesame's advantage is that the voice is embedded in a working product with memory, connectors to Gmail, Calendar and Drive, its own computer for browsing and acting, MCP support over OAuth, and, since October, the ability to orchestrate other agents. None of that is a speech-model feature; it is an assistant, and the speech is how you talk to it. Ant's advantage is that everything is inspectable. Every number in the model card can be traced to a test set, every task has a runnable example, and the pipeline is yours to modify.

The asymmetry that decides most evaluations is simpler than either list: with Ming-3.0-tts-tiny you can generate the same sentence a hundred times and listen. With Sesame you cannot generate anything at all, so there is no experiment to run — only an opinion about an app.

What Ant publishes, and what that is worth

The temptation with a matchup like this is to fill the gap with Ant's benchmark tables, and they are extensive: Chinese Seed-TTS at 0.820% word error and 0.781 speaker similarity, English at 1.666% and 0.728, German 0.325%, Spanish 0.971%, NVBench paralinguistic results leading all six single-event categories, dialect suites at 86.09% on WSC-Eval-TTS easy, emotion at 74.00% on CV3, voice design at 75.37% on InstructTTSEval-ZH.

Every one of those is Ant measuring Ant, on Ant's harness, with Ant's reference clips and scoring code, and none of it has been reproduced outside the company. That does not make the numbers wrong. It makes them vendor-reported, and it means they cannot be set against a Sesame figure because there is no Sesame figure to set them against.

The one result in the suite that behaves like a real comparison is the uncomfortable one: on Korean, Ming-3.0-tts-tiny posts 6.000% character error while its own 0.5B sibling posts 1.728%. The model card lists Korean among the known weaknesses and offers no explanation. A plausible one — sparse routing giving a lightly-trained language less capacity than a dense model of similar size would — is a hypothesis, not a finding. The operational reading is the safe one: if Korean is in scope, test it before you trust it.

Is there anything to learn from Sesame's model without the model?

Yes, and it is the most useful thing in this comparison after the licence difference: Sesame publishes a benchmark. On August 28, 2026 it released TurnBench, an open benchmark for turn-taking in spoken dialogue — 30 hours of triple-annotated dyadic speech, a scoring protocol drawn from conversation analysis, a public leaderboard, and results for 14 systems, built with Carnegie Mellon, Mundo AI and Oto.

Turn-taking is exactly the quality that a preference Elo board and a word-error table both fail to capture, and it is the quality Sesame's product is famous for. If you are evaluating Ming-3.0-tts-tiny for anything conversational, TurnBench is the instrument to borrow even though Sesame's own app is not on it: score your candidate on interruption handling and backchannel timing instead of guessing from a similarity metric.

That is the shape of a fair comparison here. Not Ming-3.0-tts-tiny versus Sesame Preview's voice, which cannot be run, but Ming-3.0-tts-tiny against a public benchmark that Sesame defined for the property both are chasing.

If you are building the thing either of them is for

The realistic 2026 architecture for a voice product is a speech layer plus a text brain that decides what to say, and these two occupy opposite ends of that stack. Ming-3.0-tts-tiny can be the speech layer, if you have a GPU and the appetite to evaluate it. Sesame Preview cannot be a layer at all; it is a competitor product you would be building against.

The text brain is the part where a router is genuinely useful rather than decorative. OrcaRouter carries more than 200 text models behind one key, at provider list price with no markup added, so a vendor's price change is your price the same day, with automatic failover across providers. To be explicit about coverage: no Ant speech model is a route here, Ming-3.0-tts-tiny included, and no Sesame model is either, CSM-1B included. What is routed is the half of the pipeline that writes the script, picks the emotion tag and decides when the agent should speak — which is also the half you can change vendors on without re-recording anything.

What to watch instead of a head-to-head

Three things would turn this into a real comparison, and none of them has happened yet.

A hosted endpoint for Ming-3.0-tts-tiny, or a third party that publishes a measured latency and error rate for it. Ant's own roadmap mentions vLLM-Omni support, which is a serving question rather than a quality one, but it is the prerequisite for anyone else scoring the model.

A number attached to Sesame's October CSM. The company says expressivity and pronunciation improved and that the difference is large over long conversations; it has published no model ID, no latency figure and no benchmark, and the only public measurement of anything adjacent remains TurnBench, which scores turn-taking rather than voice quality.

And a Korean result from anyone other than Ant. The 6.000% figure is the clearest defect in the published record, and a second team reproducing or contradicting it would tell you more about this checkpoint than any Elo score could.

Until then the honest verdict is unglamorous: one of these you can download, audit and ship, with a licence that permits it and a benchmark suite that only its author has run. The other you can talk to on your phone, and admire, and not build on.